lib/librte_eal/common/eal_common_proc.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2016-2018 Intel Corporation
   3  */
   4
   5 #include <dirent.h>
   6 #include <errno.h>
   7 #include <fcntl.h>
   8 #include <fnmatch.h>
   9 #include <inttypes.h>
  10 #include <libgen.h>
  11 #include <limits.h>
  12 #include <pthread.h>
  13 #include <stdio.h>
  14 #include <stdlib.h>
  15 #include <string.h>
  16 #include <sys/file.h>
  17 #include <sys/time.h>
  18 #include <sys/types.h>
  19 #include <sys/socket.h>
  20 #include <sys/un.h>
  21 #include <unistd.h>
  22
  23 #include <rte_common.h>
  24 #include <rte_cycles.h>
  25 #include <rte_eal.h>
  26 #include <rte_errno.h>
  27 #include <rte_lcore.h>
  28 #include <rte_log.h>
  29 #include <rte_tailq.h>
  30
  31 #include "eal_private.h"
  32 #include "eal_filesystem.h"
  33 #include "eal_internal_cfg.h"
  34
  35 static int mp_fd = -1;
  36 static char mp_filter[PATH_MAX];   /* Filter for secondary process sockets */
  37 static char mp_dir_path[PATH_MAX]; /* The directory path for all mp sockets */
  38 static pthread_mutex_t mp_mutex_action = PTHREAD_MUTEX_INITIALIZER;
  39
  40 struct action_entry {
  41         TAILQ_ENTRY(action_entry) next;
  42         char action_name[RTE_MP_MAX_NAME_LEN];
  43         rte_mp_t action;
  44 };
  45
  46 /** Double linked list of actions. */
  47 TAILQ_HEAD(action_entry_list, action_entry);
  48
  49 static struct action_entry_list action_entry_list =
  50         TAILQ_HEAD_INITIALIZER(action_entry_list);
  51
  52 enum mp_type {
  53         MP_MSG, /* Share message with peers, will not block */
  54         MP_REQ, /* Request for information, Will block for a reply */
  55         MP_REP, /* Response to previously-received request */
  56         MP_IGN, /* Response telling requester to ignore this response */
  57 };
  58
  59 struct mp_msg_internal {
  60         int type;
  61         struct rte_mp_msg msg;
  62 };
  63
  64 struct async_request_param {
  65         rte_mp_async_reply_t clb;
  66         struct rte_mp_reply user_reply;
  67         struct timespec end;
  68         int n_responses_processed;
  69 };
  70
  71 struct pending_request {
  72         TAILQ_ENTRY(pending_request) next;
  73         enum {
  74                 REQUEST_TYPE_SYNC,
  75                 REQUEST_TYPE_ASYNC
  76         } type;
  77         char dst[PATH_MAX];
  78         struct rte_mp_msg *request;
  79         struct rte_mp_msg *reply;
  80         int reply_received;
  81         RTE_STD_C11
  82         union {
  83                 struct {
  84                         struct async_request_param *param;
  85                 } async;
  86                 struct {
  87                         pthread_cond_t cond;
  88                 } sync;
  89         };
  90 };
  91
  92 TAILQ_HEAD(pending_request_list, pending_request);
  93
  94 static struct {
  95         struct pending_request_list requests;
  96         pthread_mutex_t lock;
  97         pthread_cond_t async_cond;
  98 } pending_requests = {
  99         .requests = TAILQ_HEAD_INITIALIZER(pending_requests.requests),
 100         .lock = PTHREAD_MUTEX_INITIALIZER,
 101         .async_cond = PTHREAD_COND_INITIALIZER
 102         /**< used in async requests only */
 103 };
 104
 105 /* forward declarations */
 106 static int
 107 mp_send(struct rte_mp_msg *msg, const char *peer, int type);
 108
 109
 110 static struct pending_request *
 111 find_pending_request(const char *dst, const char *act_name)
 112 {
 113         struct pending_request *r;
 114
 115         TAILQ_FOREACH(r, &pending_requests.requests, next) {
 116                 if (!strcmp(r->dst, dst) &&
 117                     !strcmp(r->request->name, act_name))
 118                         break;
 119         }
 120
 121         return r;
 122 }
 123
 124 static void
 125 create_socket_path(const char *name, char *buf, int len)
 126 {
 127         const char *prefix = eal_mp_socket_path();
 128
 129         if (strlen(name) > 0)
 130                 snprintf(buf, len, "%s_%s", prefix, name);
 131         else
 132                 strlcpy(buf, prefix, len);
 133 }
 134
 135 int
 136 rte_eal_primary_proc_alive(const char *config_file_path)
 137 {
 138         int config_fd;
 139
 140         if (config_file_path)
 141                 config_fd = open(config_file_path, O_RDONLY);
 142         else {
 143                 const char *path;
 144
 145                 path = eal_runtime_config_path();
 146                 config_fd = open(path, O_RDONLY);
 147         }
 148         if (config_fd < 0)
 149                 return 0;
 150
 151         int ret = lockf(config_fd, F_TEST, 0);
 152         close(config_fd);
 153
 154         return !!ret;
 155 }
 156
 157 static struct action_entry *
 158 find_action_entry_by_name(const char *name)
 159 {
 160         struct action_entry *entry;
 161
 162         TAILQ_FOREACH(entry, &action_entry_list, next) {
 163                 if (strncmp(entry->action_name, name, RTE_MP_MAX_NAME_LEN) == 0)
 164                         break;
 165         }
 166
 167         return entry;
 168 }
 169
 170 static int
 171 validate_action_name(const char *name)
 172 {
 173         if (name == NULL) {
 174                 RTE_LOG(ERR, EAL, "Action name cannot be NULL\n");
 175                 rte_errno = EINVAL;
 176                 return -1;
 177         }
 178         if (strnlen(name, RTE_MP_MAX_NAME_LEN) == 0) {
 179                 RTE_LOG(ERR, EAL, "Length of action name is zero\n");
 180                 rte_errno = EINVAL;
 181                 return -1;
 182         }
 183         if (strnlen(name, RTE_MP_MAX_NAME_LEN) == RTE_MP_MAX_NAME_LEN) {
 184                 rte_errno = E2BIG;
 185                 return -1;
 186         }
 187         return 0;
 188 }
 189
 190 int __rte_experimental
 191 rte_mp_action_register(const char *name, rte_mp_t action)
 192 {
 193         struct action_entry *entry;
 194
 195         if (validate_action_name(name))
 196                 return -1;
 197
 198         entry = malloc(sizeof(struct action_entry));
 199         if (entry == NULL) {
 200                 rte_errno = ENOMEM;
 201                 return -1;
 202         }
 203         strlcpy(entry->action_name, name, sizeof(entry->action_name));
 204         entry->action = action;
 205
 206         pthread_mutex_lock(&mp_mutex_action);
 207         if (find_action_entry_by_name(name) != NULL) {
 208                 pthread_mutex_unlock(&mp_mutex_action);
 209                 rte_errno = EEXIST;
 210                 free(entry);
 211                 return -1;
 212         }
 213         TAILQ_INSERT_TAIL(&action_entry_list, entry, next);
 214         pthread_mutex_unlock(&mp_mutex_action);
 215         return 0;
 216 }
 217
 218 void __rte_experimental
 219 rte_mp_action_unregister(const char *name)
 220 {
 221         struct action_entry *entry;
 222
 223         if (validate_action_name(name))
 224                 return;
 225
 226         pthread_mutex_lock(&mp_mutex_action);
 227         entry = find_action_entry_by_name(name);
 228         if (entry == NULL) {
 229                 pthread_mutex_unlock(&mp_mutex_action);
 230                 return;
 231         }
 232         TAILQ_REMOVE(&action_entry_list, entry, next);
 233         pthread_mutex_unlock(&mp_mutex_action);
 234         free(entry);
 235 }
 236
 237 static int
 238 read_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
 239 {
 240         int msglen;
 241         struct iovec iov;
 242         struct msghdr msgh;
 243         char control[CMSG_SPACE(sizeof(m->msg.fds))];
 244         struct cmsghdr *cmsg;
 245         int buflen = sizeof(*m) - sizeof(m->msg.fds);
 246
 247         memset(&msgh, 0, sizeof(msgh));
 248         iov.iov_base = m;
 249         iov.iov_len  = buflen;
 250
 251         msgh.msg_name = s;
 252         msgh.msg_namelen = sizeof(*s);
 253         msgh.msg_iov = &iov;
 254         msgh.msg_iovlen = 1;
 255         msgh.msg_control = control;
 256         msgh.msg_controllen = sizeof(control);
 257
 258         msglen = recvmsg(mp_fd, &msgh, 0);
 259         if (msglen < 0) {
 260                 RTE_LOG(ERR, EAL, "recvmsg failed, %s\n", strerror(errno));
 261                 return -1;
 262         }
 263
 264         if (msglen != buflen || (msgh.msg_flags & (MSG_TRUNC | MSG_CTRUNC))) {
 265                 RTE_LOG(ERR, EAL, "truncted msg\n");
 266                 return -1;
 267         }
 268
 269         /* read auxiliary FDs if any */
 270         for (cmsg = CMSG_FIRSTHDR(&msgh); cmsg != NULL;
 271                 cmsg = CMSG_NXTHDR(&msgh, cmsg)) {
 272                 if ((cmsg->cmsg_level == SOL_SOCKET) &&
 273                         (cmsg->cmsg_type == SCM_RIGHTS)) {
 274                         memcpy(m->msg.fds, CMSG_DATA(cmsg), sizeof(m->msg.fds));
 275                         break;
 276                 }
 277         }
 278
 279         return 0;
 280 }
 281
 282 static void
 283 process_msg(struct mp_msg_internal *m, struct sockaddr_un *s)
 284 {
 285         struct pending_request *pending_req;
 286         struct action_entry *entry;
 287         struct rte_mp_msg *msg = &m->msg;
 288         rte_mp_t action = NULL;
 289
 290         RTE_LOG(DEBUG, EAL, "msg: %s\n", msg->name);
 291
 292         if (m->type == MP_REP || m->type == MP_IGN) {
 293                 pthread_mutex_lock(&pending_requests.lock);
 294                 pending_req = find_pending_request(s->sun_path, msg->name);
 295                 if (pending_req) {
 296                         memcpy(pending_req->reply, msg, sizeof(*msg));
 297                         /* -1 indicates that we've been asked to ignore */
 298                         pending_req->reply_received =
 299                                 m->type == MP_REP ? 1 : -1;
 300
 301                         if (pending_req->type == REQUEST_TYPE_SYNC)
 302                                 pthread_cond_signal(&pending_req->sync.cond);
 303                         else if (pending_req->type == REQUEST_TYPE_ASYNC)
 304                                 pthread_cond_signal(
 305                                         &pending_requests.async_cond);
 306                 } else
 307                         RTE_LOG(ERR, EAL, "Drop mp reply: %s\n", msg->name);
 308                 pthread_mutex_unlock(&pending_requests.lock);
 309                 return;
 310         }
 311
 312         pthread_mutex_lock(&mp_mutex_action);
 313         entry = find_action_entry_by_name(msg->name);
 314         if (entry != NULL)
 315                 action = entry->action;
 316         pthread_mutex_unlock(&mp_mutex_action);
 317
 318         if (!action) {
 319                 if (m->type == MP_REQ && !internal_config.init_complete) {
 320                         /* if this is a request, and init is not yet complete,
 321                          * and callback wasn't registered, we should tell the
 322                          * requester to ignore our existence because we're not
 323                          * yet ready to process this request.
 324                          */
 325                         struct rte_mp_msg dummy;
 326
 327                         memset(&dummy, 0, sizeof(dummy));
 328                         strlcpy(dummy.name, msg->name, sizeof(dummy.name));
 329                         mp_send(&dummy, s->sun_path, MP_IGN);
 330                 } else {
 331                         RTE_LOG(ERR, EAL, "Cannot find action: %s\n",
 332                                 msg->name);
 333                 }
 334         } else if (action(msg, s->sun_path) < 0) {
 335                 RTE_LOG(ERR, EAL, "Fail to handle message: %s\n", msg->name);
 336         }
 337 }
 338
 339 static void *
 340 mp_handle(void *arg __rte_unused)
 341 {
 342         struct mp_msg_internal msg;
 343         struct sockaddr_un sa;
 344
 345         while (1) {
 346                 if (read_msg(&msg, &sa) == 0)
 347                         process_msg(&msg, &sa);
 348         }
 349
 350         return NULL;
 351 }
 352
 353 static int
 354 timespec_cmp(const struct timespec *a, const struct timespec *b)
 355 {
 356         if (a->tv_sec < b->tv_sec)
 357                 return -1;
 358         if (a->tv_sec > b->tv_sec)
 359                 return 1;
 360         if (a->tv_nsec < b->tv_nsec)
 361                 return -1;
 362         if (a->tv_nsec > b->tv_nsec)
 363                 return 1;
 364         return 0;
 365 }
 366
 367 enum async_action {
 368         ACTION_NONE, /**< don't do anything */
 369         ACTION_FREE, /**< free the action entry, but don't trigger callback */
 370         ACTION_TRIGGER /**< trigger callback, then free action entry */
 371 };
 372
 373 static enum async_action
 374 process_async_request(struct pending_request *sr, const struct timespec *now)
 375 {
 376         struct async_request_param *param;
 377         struct rte_mp_reply *reply;
 378         bool timeout, received, last_msg;
 379
 380         param = sr->async.param;
 381         reply = &param->user_reply;
 382
 383         /* did we timeout? */
 384         timeout = timespec_cmp(&param->end, now) <= 0;
 385
 386         /* did we receive a response? */
 387         received = sr->reply_received != 0;
 388
 389         /* if we didn't time out, and we didn't receive a response, ignore */
 390         if (!timeout && !received)
 391                 return ACTION_NONE;
 392
 393         /* if we received a response, adjust relevant data and copy mesasge. */
 394         if (sr->reply_received == 1 && sr->reply) {
 395                 struct rte_mp_msg *msg, *user_msgs, *tmp;
 396
 397                 msg = sr->reply;
 398                 user_msgs = reply->msgs;
 399
 400                 tmp = realloc(user_msgs, sizeof(*msg) *
 401                                 (reply->nb_received + 1));
 402                 if (!tmp) {
 403                         RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 404                                 sr->dst, sr->request->name);
 405                         /* this entry is going to be removed and its message
 406                          * dropped, but we don't want to leak memory, so
 407                          * continue.
 408                          */
 409                 } else {
 410                         user_msgs = tmp;
 411                         reply->msgs = user_msgs;
 412                         memcpy(&user_msgs[reply->nb_received],
 413                                         msg, sizeof(*msg));
 414                         reply->nb_received++;
 415                 }
 416
 417                 /* mark this request as processed */
 418                 param->n_responses_processed++;
 419         } else if (sr->reply_received == -1) {
 420                 /* we were asked to ignore this process */
 421                 reply->nb_sent--;
 422         } else if (timeout) {
 423                 /* count it as processed response, but don't increment
 424                  * nb_received.
 425                  */
 426                 param->n_responses_processed++;
 427         }
 428
 429         free(sr->reply);
 430
 431         last_msg = param->n_responses_processed == reply->nb_sent;
 432
 433         return last_msg ? ACTION_TRIGGER : ACTION_FREE;
 434 }
 435
 436 static void
 437 trigger_async_action(struct pending_request *sr)
 438 {
 439         struct async_request_param *param;
 440         struct rte_mp_reply *reply;
 441
 442         param = sr->async.param;
 443         reply = &param->user_reply;
 444
 445         param->clb(sr->request, reply);
 446
 447         /* clean up */
 448         free(sr->async.param->user_reply.msgs);
 449         free(sr->async.param);
 450         free(sr->request);
 451 }
 452
 453 static struct pending_request *
 454 check_trigger(struct timespec *ts)
 455 {
 456         struct pending_request *next, *cur, *trigger = NULL;
 457
 458         TAILQ_FOREACH_SAFE(cur, &pending_requests.requests, next, next) {
 459                 enum async_action action;
 460                 if (cur->type != REQUEST_TYPE_ASYNC)
 461                         continue;
 462
 463                 action = process_async_request(cur, ts);
 464                 if (action == ACTION_FREE) {
 465                         TAILQ_REMOVE(&pending_requests.requests, cur, next);
 466                         free(cur);
 467                 } else if (action == ACTION_TRIGGER) {
 468                         TAILQ_REMOVE(&pending_requests.requests, cur, next);
 469                         trigger = cur;
 470                         break;
 471                 }
 472         }
 473         return trigger;
 474 }
 475
 476 static void
 477 wait_for_async_messages(void)
 478 {
 479         struct pending_request *sr;
 480         struct timespec timeout;
 481         bool timedwait = false;
 482         bool nowait = false;
 483         int ret;
 484
 485         /* scan through the list and see if there are any timeouts that
 486          * are earlier than our current timeout.
 487          */
 488         TAILQ_FOREACH(sr, &pending_requests.requests, next) {
 489                 if (sr->type != REQUEST_TYPE_ASYNC)
 490                         continue;
 491                 if (!timedwait || timespec_cmp(&sr->async.param->end,
 492                                 &timeout) < 0) {
 493                         memcpy(&timeout, &sr->async.param->end,
 494                                 sizeof(timeout));
 495                         timedwait = true;
 496                 }
 497
 498                 /* sometimes, we don't even wait */
 499                 if (sr->reply_received) {
 500                         nowait = true;
 501                         break;
 502                 }
 503         }
 504
 505         if (nowait)
 506                 return;
 507
 508         do {
 509                 ret = timedwait ?
 510                         pthread_cond_timedwait(
 511                                 &pending_requests.async_cond,
 512                                 &pending_requests.lock,
 513                                 &timeout) :
 514                         pthread_cond_wait(
 515                                 &pending_requests.async_cond,
 516                                 &pending_requests.lock);
 517         } while (ret != 0 && ret != ETIMEDOUT);
 518
 519         /* we've been woken up or timed out */
 520 }
 521
 522 static void *
 523 async_reply_handle(void *arg __rte_unused)
 524 {
 525         struct timeval now;
 526         struct timespec ts_now;
 527         while (1) {
 528                 struct pending_request *trigger = NULL;
 529
 530                 pthread_mutex_lock(&pending_requests.lock);
 531
 532                 /* we exit this function holding the lock */
 533                 wait_for_async_messages();
 534
 535                 if (gettimeofday(&now, NULL) < 0) {
 536                         pthread_mutex_unlock(&pending_requests.lock);
 537                         RTE_LOG(ERR, EAL, "Cannot get current time\n");
 538                         break;
 539                 }
 540                 ts_now.tv_nsec = now.tv_usec * 1000;
 541                 ts_now.tv_sec = now.tv_sec;
 542
 543                 do {
 544                         trigger = check_trigger(&ts_now);
 545                         /* unlock request list */
 546                         pthread_mutex_unlock(&pending_requests.lock);
 547
 548                         if (trigger) {
 549                                 trigger_async_action(trigger);
 550                                 free(trigger);
 551
 552                                 /* we've triggered a callback, but there may be
 553                                  * more, so lock the list and check again.
 554                                  */
 555                                 pthread_mutex_lock(&pending_requests.lock);
 556                         }
 557                 } while (trigger);
 558         }
 559
 560         RTE_LOG(ERR, EAL, "ERROR: asynchronous requests disabled\n");
 561
 562         return NULL;
 563 }
 564
 565 static int
 566 open_socket_fd(void)
 567 {
 568         char peer_name[PATH_MAX] = {0};
 569         struct sockaddr_un un;
 570
 571         if (rte_eal_process_type() == RTE_PROC_SECONDARY)
 572                 snprintf(peer_name, sizeof(peer_name),
 573                                 "%d_%"PRIx64, getpid(), rte_rdtsc());
 574
 575         mp_fd = socket(AF_UNIX, SOCK_DGRAM, 0);
 576         if (mp_fd < 0) {
 577                 RTE_LOG(ERR, EAL, "failed to create unix socket\n");
 578                 return -1;
 579         }
 580
 581         memset(&un, 0, sizeof(un));
 582         un.sun_family = AF_UNIX;
 583
 584         create_socket_path(peer_name, un.sun_path, sizeof(un.sun_path));
 585
 586         unlink(un.sun_path); /* May still exist since last run */
 587
 588         if (bind(mp_fd, (struct sockaddr *)&un, sizeof(un)) < 0) {
 589                 RTE_LOG(ERR, EAL, "failed to bind %s: %s\n",
 590                         un.sun_path, strerror(errno));
 591                 close(mp_fd);
 592                 return -1;
 593         }
 594
 595         RTE_LOG(INFO, EAL, "Multi-process socket %s\n", un.sun_path);
 596         return mp_fd;
 597 }
 598
 599 static int
 600 unlink_sockets(const char *filter)
 601 {
 602         int dir_fd;
 603         DIR *mp_dir;
 604         struct dirent *ent;
 605
 606         mp_dir = opendir(mp_dir_path);
 607         if (!mp_dir) {
 608                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
 609                 return -1;
 610         }
 611         dir_fd = dirfd(mp_dir);
 612
 613         while ((ent = readdir(mp_dir))) {
 614                 if (fnmatch(filter, ent->d_name, 0) == 0)
 615                         unlinkat(dir_fd, ent->d_name, 0);
 616         }
 617
 618         closedir(mp_dir);
 619         return 0;
 620 }
 621
 622 int
 623 rte_mp_channel_init(void)
 624 {
 625         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 626         char path[PATH_MAX];
 627         int dir_fd;
 628         pthread_t mp_handle_tid, async_reply_handle_tid;
 629
 630         /* create filter path */
 631         create_socket_path("*", path, sizeof(path));
 632         strlcpy(mp_filter, basename(path), sizeof(mp_filter));
 633
 634         /* path may have been modified, so recreate it */
 635         create_socket_path("*", path, sizeof(path));
 636         strlcpy(mp_dir_path, dirname(path), sizeof(mp_dir_path));
 637
 638         /* lock the directory */
 639         dir_fd = open(mp_dir_path, O_RDONLY);
 640         if (dir_fd < 0) {
 641                 RTE_LOG(ERR, EAL, "failed to open %s: %s\n",
 642                         mp_dir_path, strerror(errno));
 643                 return -1;
 644         }
 645
 646         if (flock(dir_fd, LOCK_EX)) {
 647                 RTE_LOG(ERR, EAL, "failed to lock %s: %s\n",
 648                         mp_dir_path, strerror(errno));
 649                 close(dir_fd);
 650                 return -1;
 651         }
 652
 653         if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
 654                         unlink_sockets(mp_filter)) {
 655                 RTE_LOG(ERR, EAL, "failed to unlink mp sockets\n");
 656                 close(dir_fd);
 657                 return -1;
 658         }
 659
 660         if (open_socket_fd() < 0) {
 661                 close(dir_fd);
 662                 return -1;
 663         }
 664
 665         if (rte_ctrl_thread_create(&mp_handle_tid, NULL, mp_handle, NULL) < 0) {
 666                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 667                         strerror(errno));
 668                 close(mp_fd);
 669                 close(dir_fd);
 670                 mp_fd = -1;
 671                 return -1;
 672         }
 673
 674         if (rte_ctrl_thread_create(&async_reply_handle_tid, NULL,
 675                         async_reply_handle, NULL) < 0) {
 676                 RTE_LOG(ERR, EAL, "failed to create mp thead: %s\n",
 677                         strerror(errno));
 678                 close(mp_fd);
 679                 close(dir_fd);
 680                 mp_fd = -1;
 681                 return -1;
 682         }
 683
 684         /* try best to set thread name */
 685         strlcpy(thread_name, "rte_mp_handle", sizeof(thread_name));
 686         rte_thread_setname(mp_handle_tid, thread_name);
 687
 688         /* try best to set thread name */
 689         strlcpy(thread_name, "rte_mp_async_handle", sizeof(thread_name));
 690         rte_thread_setname(async_reply_handle_tid, thread_name);
 691
 692         /* unlock the directory */
 693         flock(dir_fd, LOCK_UN);
 694         close(dir_fd);
 695
 696         return 0;
 697 }
 698
 699 /**
 700  * Return -1, as fail to send message and it's caused by the local side.
 701  * Return 0, as fail to send message and it's caused by the remote side.
 702  * Return 1, as succeed to send message.
 703  *
 704  */
 705 static int
 706 send_msg(const char *dst_path, struct rte_mp_msg *msg, int type)
 707 {
 708         int snd;
 709         struct iovec iov;
 710         struct msghdr msgh;
 711         struct cmsghdr *cmsg;
 712         struct sockaddr_un dst;
 713         struct mp_msg_internal m;
 714         int fd_size = msg->num_fds * sizeof(int);
 715         char control[CMSG_SPACE(fd_size)];
 716
 717         m.type = type;
 718         memcpy(&m.msg, msg, sizeof(*msg));
 719
 720         memset(&dst, 0, sizeof(dst));
 721         dst.sun_family = AF_UNIX;
 722         strlcpy(dst.sun_path, dst_path, sizeof(dst.sun_path));
 723
 724         memset(&msgh, 0, sizeof(msgh));
 725         memset(control, 0, sizeof(control));
 726
 727         iov.iov_base = &m;
 728         iov.iov_len = sizeof(m) - sizeof(msg->fds);
 729
 730         msgh.msg_name = &dst;
 731         msgh.msg_namelen = sizeof(dst);
 732         msgh.msg_iov = &iov;
 733         msgh.msg_iovlen = 1;
 734         msgh.msg_control = control;
 735         msgh.msg_controllen = sizeof(control);
 736
 737         cmsg = CMSG_FIRSTHDR(&msgh);
 738         cmsg->cmsg_len = CMSG_LEN(fd_size);
 739         cmsg->cmsg_level = SOL_SOCKET;
 740         cmsg->cmsg_type = SCM_RIGHTS;
 741         memcpy(CMSG_DATA(cmsg), msg->fds, fd_size);
 742
 743         do {
 744                 snd = sendmsg(mp_fd, &msgh, 0);
 745         } while (snd < 0 && errno == EINTR);
 746
 747         if (snd < 0) {
 748                 rte_errno = errno;
 749                 /* Check if it caused by peer process exits */
 750                 if (errno == ECONNREFUSED &&
 751                                 rte_eal_process_type() == RTE_PROC_PRIMARY) {
 752                         unlink(dst_path);
 753                         return 0;
 754                 }
 755                 if (errno == ENOBUFS) {
 756                         RTE_LOG(ERR, EAL, "Peer cannot receive message %s\n",
 757                                 dst_path);
 758                         return 0;
 759                 }
 760                 RTE_LOG(ERR, EAL, "failed to send to (%s) due to %s\n",
 761                         dst_path, strerror(errno));
 762                 return -1;
 763         }
 764
 765         return 1;
 766 }
 767
 768 static int
 769 mp_send(struct rte_mp_msg *msg, const char *peer, int type)
 770 {
 771         int dir_fd, ret = 0;
 772         DIR *mp_dir;
 773         struct dirent *ent;
 774
 775         if (!peer && (rte_eal_process_type() == RTE_PROC_SECONDARY))
 776                 peer = eal_mp_socket_path();
 777
 778         if (peer) {
 779                 if (send_msg(peer, msg, type) < 0)
 780                         return -1;
 781                 else
 782                         return 0;
 783         }
 784
 785         /* broadcast to all secondary processes */
 786         mp_dir = opendir(mp_dir_path);
 787         if (!mp_dir) {
 788                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n",
 789                                 mp_dir_path);
 790                 rte_errno = errno;
 791                 return -1;
 792         }
 793
 794         dir_fd = dirfd(mp_dir);
 795         /* lock the directory to prevent processes spinning up while we send */
 796         if (flock(dir_fd, LOCK_EX)) {
 797                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
 798                         mp_dir_path);
 799                 rte_errno = errno;
 800                 closedir(mp_dir);
 801                 return -1;
 802         }
 803
 804         while ((ent = readdir(mp_dir))) {
 805                 char path[PATH_MAX];
 806
 807                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
 808                         continue;
 809
 810                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
 811                          ent->d_name);
 812                 if (send_msg(path, msg, type) < 0)
 813                         ret = -1;
 814         }
 815         /* unlock the dir */
 816         flock(dir_fd, LOCK_UN);
 817
 818         /* dir_fd automatically closed on closedir */
 819         closedir(mp_dir);
 820         return ret;
 821 }
 822
 823 static bool
 824 check_input(const struct rte_mp_msg *msg)
 825 {
 826         if (msg == NULL) {
 827                 RTE_LOG(ERR, EAL, "Msg cannot be NULL\n");
 828                 rte_errno = EINVAL;
 829                 return false;
 830         }
 831
 832         if (validate_action_name(msg->name))
 833                 return false;
 834
 835         if (msg->len_param > RTE_MP_MAX_PARAM_LEN) {
 836                 RTE_LOG(ERR, EAL, "Message data is too long\n");
 837                 rte_errno = E2BIG;
 838                 return false;
 839         }
 840
 841         if (msg->num_fds > RTE_MP_MAX_FD_NUM) {
 842                 RTE_LOG(ERR, EAL, "Cannot send more than %d FDs\n",
 843                         RTE_MP_MAX_FD_NUM);
 844                 rte_errno = E2BIG;
 845                 return false;
 846         }
 847
 848         return true;
 849 }
 850
 851 int __rte_experimental
 852 rte_mp_sendmsg(struct rte_mp_msg *msg)
 853 {
 854         if (!check_input(msg))
 855                 return -1;
 856
 857         RTE_LOG(DEBUG, EAL, "sendmsg: %s\n", msg->name);
 858         return mp_send(msg, NULL, MP_MSG);
 859 }
 860
 861 static int
 862 mp_request_async(const char *dst, struct rte_mp_msg *req,
 863                 struct async_request_param *param)
 864 {
 865         struct rte_mp_msg *reply_msg;
 866         struct pending_request *pending_req, *exist;
 867         int ret;
 868
 869         pending_req = calloc(1, sizeof(*pending_req));
 870         reply_msg = calloc(1, sizeof(*reply_msg));
 871         if (pending_req == NULL || reply_msg == NULL) {
 872                 RTE_LOG(ERR, EAL, "Could not allocate space for sync request\n");
 873                 rte_errno = ENOMEM;
 874                 ret = -1;
 875                 goto fail;
 876         }
 877
 878         pending_req->type = REQUEST_TYPE_ASYNC;
 879         strlcpy(pending_req->dst, dst, sizeof(pending_req->dst));
 880         strcpy(pending_req->dst, dst);
 881         pending_req->request = req;
 882         pending_req->reply = reply_msg;
 883         pending_req->async.param = param;
 884
 885         /* queue already locked by caller */
 886
 887         exist = find_pending_request(dst, req->name);
 888         if (exist) {
 889                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 890                 rte_errno = EEXIST;
 891                 ret = -1;
 892                 goto fail;
 893         }
 894
 895         ret = send_msg(dst, req, MP_REQ);
 896         if (ret < 0) {
 897                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 898                         dst, req->name);
 899                 ret = -1;
 900                 goto fail;
 901         } else if (ret == 0) {
 902                 ret = 0;
 903                 goto fail;
 904         }
 905         TAILQ_INSERT_TAIL(&pending_requests.requests, pending_req, next);
 906
 907         param->user_reply.nb_sent++;
 908
 909         return 0;
 910 fail:
 911         free(pending_req);
 912         free(reply_msg);
 913         return ret;
 914 }
 915
 916 static int
 917 mp_request_sync(const char *dst, struct rte_mp_msg *req,
 918                struct rte_mp_reply *reply, const struct timespec *ts)
 919 {
 920         int ret;
 921         struct rte_mp_msg msg, *tmp;
 922         struct pending_request pending_req, *exist;
 923
 924         pending_req.type = REQUEST_TYPE_SYNC;
 925         pending_req.reply_received = 0;
 926         strlcpy(pending_req.dst, dst, sizeof(pending_req.dst));
 927         pending_req.request = req;
 928         pending_req.reply = &msg;
 929         pthread_cond_init(&pending_req.sync.cond, NULL);
 930
 931         exist = find_pending_request(dst, req->name);
 932         if (exist) {
 933                 RTE_LOG(ERR, EAL, "A pending request %s:%s\n", dst, req->name);
 934                 rte_errno = EEXIST;
 935                 return -1;
 936         }
 937
 938         ret = send_msg(dst, req, MP_REQ);
 939         if (ret < 0) {
 940                 RTE_LOG(ERR, EAL, "Fail to send request %s:%s\n",
 941                         dst, req->name);
 942                 return -1;
 943         } else if (ret == 0)
 944                 return 0;
 945
 946         TAILQ_INSERT_TAIL(&pending_requests.requests, &pending_req, next);
 947
 948         reply->nb_sent++;
 949
 950         do {
 951                 ret = pthread_cond_timedwait(&pending_req.sync.cond,
 952                                 &pending_requests.lock, ts);
 953         } while (ret != 0 && ret != ETIMEDOUT);
 954
 955         TAILQ_REMOVE(&pending_requests.requests, &pending_req, next);
 956
 957         if (pending_req.reply_received == 0) {
 958                 RTE_LOG(ERR, EAL, "Fail to recv reply for request %s:%s\n",
 959                         dst, req->name);
 960                 rte_errno = ETIMEDOUT;
 961                 return -1;
 962         }
 963         if (pending_req.reply_received == -1) {
 964                 RTE_LOG(DEBUG, EAL, "Asked to ignore response\n");
 965                 /* not receiving this message is not an error, so decrement
 966                  * number of sent messages
 967                  */
 968                 reply->nb_sent--;
 969                 return 0;
 970         }
 971
 972         tmp = realloc(reply->msgs, sizeof(msg) * (reply->nb_received + 1));
 973         if (!tmp) {
 974                 RTE_LOG(ERR, EAL, "Fail to alloc reply for request %s:%s\n",
 975                         dst, req->name);
 976                 rte_errno = ENOMEM;
 977                 return -1;
 978         }
 979         memcpy(&tmp[reply->nb_received], &msg, sizeof(msg));
 980         reply->msgs = tmp;
 981         reply->nb_received++;
 982         return 0;
 983 }
 984
 985 int __rte_experimental
 986 rte_mp_request_sync(struct rte_mp_msg *req, struct rte_mp_reply *reply,
 987                 const struct timespec *ts)
 988 {
 989         int dir_fd, ret = 0;
 990         DIR *mp_dir;
 991         struct dirent *ent;
 992         struct timeval now;
 993         struct timespec end;
 994
 995         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
 996
 997         if (check_input(req) == false)
 998                 return -1;
 999         if (gettimeofday(&now, NULL) < 0) {
1000                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
1001                 rte_errno = errno;
1002                 return -1;
1003         }
1004
1005         end.tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
1006         end.tv_sec = now.tv_sec + ts->tv_sec +
1007                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1008
1009         reply->nb_sent = 0;
1010         reply->nb_received = 0;
1011         reply->msgs = NULL;
1012
1013         /* for secondary process, send request to the primary process only */
1014         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1015                 pthread_mutex_lock(&pending_requests.lock);
1016                 ret = mp_request_sync(eal_mp_socket_path(), req, reply, &end);
1017                 pthread_mutex_unlock(&pending_requests.lock);
1018                 return ret;
1019         }
1020
1021         /* for primary process, broadcast request, and collect reply 1 by 1 */
1022         mp_dir = opendir(mp_dir_path);
1023         if (!mp_dir) {
1024                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1025                 rte_errno = errno;
1026                 return -1;
1027         }
1028
1029         dir_fd = dirfd(mp_dir);
1030         /* lock the directory to prevent processes spinning up while we send */
1031         if (flock(dir_fd, LOCK_EX)) {
1032                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1033                         mp_dir_path);
1034                 closedir(mp_dir);
1035                 rte_errno = errno;
1036                 return -1;
1037         }
1038
1039         pthread_mutex_lock(&pending_requests.lock);
1040         while ((ent = readdir(mp_dir))) {
1041                 char path[PATH_MAX];
1042
1043                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1044                         continue;
1045
1046                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1047                          ent->d_name);
1048
1049                 /* unlocks the mutex while waiting for response,
1050                  * locks on receive
1051                  */
1052                 if (mp_request_sync(path, req, reply, &end))
1053                         ret = -1;
1054         }
1055         pthread_mutex_unlock(&pending_requests.lock);
1056         /* unlock the directory */
1057         flock(dir_fd, LOCK_UN);
1058
1059         /* dir_fd automatically closed on closedir */
1060         closedir(mp_dir);
1061         return ret;
1062 }
1063
1064 int __rte_experimental
1065 rte_mp_request_async(struct rte_mp_msg *req, const struct timespec *ts,
1066                 rte_mp_async_reply_t clb)
1067 {
1068         struct rte_mp_msg *copy;
1069         struct pending_request *dummy;
1070         struct async_request_param *param;
1071         struct rte_mp_reply *reply;
1072         int dir_fd, ret = 0;
1073         DIR *mp_dir;
1074         struct dirent *ent;
1075         struct timeval now;
1076         struct timespec *end;
1077         bool dummy_used = false;
1078
1079         RTE_LOG(DEBUG, EAL, "request: %s\n", req->name);
1080
1081         if (check_input(req) == false)
1082                 return -1;
1083         if (gettimeofday(&now, NULL) < 0) {
1084                 RTE_LOG(ERR, EAL, "Faile to get current time\n");
1085                 rte_errno = errno;
1086                 return -1;
1087         }
1088         copy = calloc(1, sizeof(*copy));
1089         dummy = calloc(1, sizeof(*dummy));
1090         param = calloc(1, sizeof(*param));
1091         if (copy == NULL || dummy == NULL || param == NULL) {
1092                 RTE_LOG(ERR, EAL, "Failed to allocate memory for async reply\n");
1093                 rte_errno = ENOMEM;
1094                 goto fail;
1095         }
1096
1097         /* copy message */
1098         memcpy(copy, req, sizeof(*copy));
1099
1100         param->n_responses_processed = 0;
1101         param->clb = clb;
1102         end = &param->end;
1103         reply = &param->user_reply;
1104
1105         end->tv_nsec = (now.tv_usec * 1000 + ts->tv_nsec) % 1000000000;
1106         end->tv_sec = now.tv_sec + ts->tv_sec +
1107                         (now.tv_usec * 1000 + ts->tv_nsec) / 1000000000;
1108         reply->nb_sent = 0;
1109         reply->nb_received = 0;
1110         reply->msgs = NULL;
1111
1112         /* we have to lock the request queue here, as we will be adding a bunch
1113          * of requests to the queue at once, and some of the replies may arrive
1114          * before we add all of the requests to the queue.
1115          */
1116         pthread_mutex_lock(&pending_requests.lock);
1117
1118         /* we have to ensure that callback gets triggered even if we don't send
1119          * anything, therefore earlier we have allocated a dummy request. fill
1120          * it, and put it on the queue if we don't send any requests.
1121          */
1122         dummy->type = REQUEST_TYPE_ASYNC;
1123         dummy->request = copy;
1124         dummy->reply = NULL;
1125         dummy->async.param = param;
1126         dummy->reply_received = 1; /* short-circuit the timeout */
1127
1128         /* for secondary process, send request to the primary process only */
1129         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1130                 ret = mp_request_async(eal_mp_socket_path(), copy, param);
1131
1132                 /* if we didn't send anything, put dummy request on the queue */
1133                 if (ret == 0 && reply->nb_sent == 0) {
1134                         TAILQ_INSERT_TAIL(&pending_requests.requests, dummy,
1135                                         next);
1136                         dummy_used = true;
1137                 }
1138
1139                 pthread_mutex_unlock(&pending_requests.lock);
1140
1141                 /* if we couldn't send anything, clean up */
1142                 if (ret != 0)
1143                         goto fail;
1144                 return 0;
1145         }
1146
1147         /* for primary process, broadcast request */
1148         mp_dir = opendir(mp_dir_path);
1149         if (!mp_dir) {
1150                 RTE_LOG(ERR, EAL, "Unable to open directory %s\n", mp_dir_path);
1151                 rte_errno = errno;
1152                 goto unlock_fail;
1153         }
1154         dir_fd = dirfd(mp_dir);
1155
1156         /* lock the directory to prevent processes spinning up while we send */
1157         if (flock(dir_fd, LOCK_EX)) {
1158                 RTE_LOG(ERR, EAL, "Unable to lock directory %s\n",
1159                         mp_dir_path);
1160                 rte_errno = errno;
1161                 goto closedir_fail;
1162         }
1163
1164         while ((ent = readdir(mp_dir))) {
1165                 char path[PATH_MAX];
1166
1167                 if (fnmatch(mp_filter, ent->d_name, 0) != 0)
1168                         continue;
1169
1170                 snprintf(path, sizeof(path), "%s/%s", mp_dir_path,
1171                          ent->d_name);
1172
1173                 if (mp_request_async(path, copy, param))
1174                         ret = -1;
1175         }
1176         /* if we didn't send anything, put dummy request on the queue */
1177         if (ret == 0 && reply->nb_sent == 0) {
1178                 TAILQ_INSERT_HEAD(&pending_requests.requests, dummy, next);
1179                 dummy_used = true;
1180         }
1181
1182         /* trigger async request thread wake up */
1183         pthread_cond_signal(&pending_requests.async_cond);
1184
1185         /* finally, unlock the queue */
1186         pthread_mutex_unlock(&pending_requests.lock);
1187
1188         /* unlock the directory */
1189         flock(dir_fd, LOCK_UN);
1190
1191         /* dir_fd automatically closed on closedir */
1192         closedir(mp_dir);
1193
1194         /* if dummy was unused, free it */
1195         if (!dummy_used)
1196                 free(dummy);
1197
1198         return ret;
1199 closedir_fail:
1200         closedir(mp_dir);
1201 unlock_fail:
1202         pthread_mutex_unlock(&pending_requests.lock);
1203 fail:
1204         free(dummy);
1205         free(param);
1206         free(copy);
1207         return -1;
1208 }
1209
1210 int __rte_experimental
1211 rte_mp_reply(struct rte_mp_msg *msg, const char *peer)
1212 {
1213         RTE_LOG(DEBUG, EAL, "reply: %s\n", msg->name);
1214
1215         if (check_input(msg) == false)
1216                 return -1;
1217
1218         if (peer == NULL) {
1219                 RTE_LOG(ERR, EAL, "peer is not specified\n");
1220                 rte_errno = EINVAL;
1221                 return -1;
1222         }
1223
1224         return mp_send(msg, peer, MP_REP);
1225 }